Browse Prior Art Database

Device Skew Management in Dual Copy DASD Environment

IP.com Disclosure Number: IPCOM000116206D
Original Publication Date: 1995-Aug-01
Included in the Prior Art Database: 2005-Mar-30
Document File: 4 page(s) / 141K

Publishing Venue

IBM

Related People

Onton, A: AUTHOR

Abstract

Device skew in many DASD installations limits the throughput of the storage subsystem because the busiest device acts as a bottleneck for the entire processor/storage complex. Many of the busiest installations also maintain two copies of all data on DASD for data availability and protection against data loss [1-4]. This method proposes a data allocation and access algorithm by which device skew in such a dual copy environment can be reduced.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 40% of the total text.

Device Skew Management in Dual Copy DASD Environment

      Device skew in many DASD installations limits the throughput of
the storage subsystem because the busiest device acts as a bottleneck
for the entire processor/storage complex.  Many of the busiest
installations also maintain two copies of all data on DASD for data
availability and protection against data loss [1-4].  This method
proposes a data allocation and access algorithm by which device skew
in such a dual copy environment can be reduced.

      In order to demonstrate the opportunity of managing device skew
in the dual copy environment, we take a conservative sample skew
(henceforth referred to as "nominal") wherein the most active device
is experiencing twice the request rate of the system average per
device.  (This example is for illustrative purposes only.)  Assume a
string of eight (8) independent devices, and arbitrarily normalize
the device activity to a peak value of ten (10).

Assume the following nominal device skew:
  Device #:  ^^1   ^2   ^3   ^4   ^5   ^6   ^7   ^8  ^= i
  Activity: ^^10   ^8   ^6   ^5   ^4   ^3   ^2   ^2  ^= "S" sup t
   (<S sub avg> = 5$$ 'requests/s')

Then define the following algorithm:
  1.  Allocate the 1st copy of the data according to existing
practice
  2.  Allocate the 2nd copy randomly to any other device
  3.  Maintain a value of the short term activity rate for each
device,
       S sub i
  4.  Order devices by descending activity
  5.  Calculate average activity rate per device, S sub avg
  6.  For  writes,  handle  the write with the shorter access time,
the
       second write can be buffered for execution at lower priority
  7.  For reads, handle up to a fraction S sub avg / S sub 1,
including
       writes, of incoming requests at the primary device, referring
the
       remainder to the secondary device
  8.  If the secondary reference read is to a higher  ranked  device,
       handle at the primary device

      The function  of the above algorithm on the "nominal" skew can
be expressed as the following, example specific, transfer matrix "R"
which converts the vector of original device activities, "S", to the
"zero-skew" vector of device activities, "S" sup 0, as follows:
  <"S" sup 0> = <left lparen <"R" over "S"> right rparen sup t> "S",
  where the simple division of the matrix "R" by the vector "S"
implies
that the first row of "R" is divided by the first element of "S",
etc.
  "R", for the "read-only" case, is constructed according to the
algorithm as follows:
  Device
  handling  j= ^^1    ^2    ^3    ^4    ^5    ^6    ^7    ^8
   request
    ^^^^(units of matrix = requests/s)
  "S" sup 0t=sum (i)=^^5   ^5.3  ^5.1  ^5.2  ^5.2  ^5.0  ^4.4  ^4.
  where, starting at the upper left  term  we  handle  5  of  the
requests  coming in to device 1 at device 1 (i.e., the fraction S su...